The propagation‐induced degradation of classification ability for echoes from a finite cylindrical shell and a from a fish school is investigated by use of the class specific method [P.M. Baggenstoss, ‘‘Class‐specific feature sets in classification,’’ IEEE Trans. Sig. Proc. 47, 3428–3432 (1999)]. The study is performed with previously displayed numerically simulated data. In prior work, degradation from additive noise was studied at a variety of signal bandwidths. Here, convolution noise caused by two‐way propagation in a shallow water environment is shown to increase the duration of non‐propagated echoes. The common propagation path between the two classes causes degradation when the classifier is based on non‐propagated training data. Ameliorating effects of using training data comprised of either a fixed range or multiple distinct ranges are shown. The performance is quantitatively displayed as either classifier operating characteristic (CLOC) curves or as the integrals of the CLOC curves. The design trade‐off space of bandwidth, signal‐to‐noise ratio (proportional to source level), and complexity of the training set is shown over a limited range of values. [Work supported by ONR.]